Gaussian Splatting

Gaussian Splatting is a very powerful representation method of 3D geometry. Once the representation is optimized for a certain scene, it can be rendered from arbitrary viewpoints. To control the number of points, Kerbl et al. use density-based pruning, splitting, and cloning. Splitting and cloning are methods to create additional Gaussians to represent the geometry more accurately. Pruning is a method to remove Gaussians that are not necessary to represent the geometry by checking if the point's opacity is below a threshold.

Fix the number of points used in the scene

In this project, we utilize the opacity-based pruning method to give the user control over the number of points that are used to model the geometry. When setting this parameter to $4,000$, the optimization times decreases significantly. And for simple ShapeNet scenes, the quality of the reconstructions does not suffer. In fact, we observed a 2% increase in PSNR, LPIPS, and SSIM when fixing the number of points to $4,000$ instead of not exerting any control over the number of parameters.

Leveraging depth maps

Instead of initializing the Gaussian point cloud randomly, this repository also offers the opportunity of leveraging depth maps to initialize the scene with a better geometry.